Intelligent prediction method and system for quality of ester oil raw material

Abstract

The invention discloses a method and a system for intelligently predicting the quality of an ester oil raw material. The method comprises the following steps: A1, acquiring characteristic data of base oil; processing parameter data in the base oil refining process are collected and preprocessed; a2, respectively extracting chemical characteristics and processing time sequence characteristics of the base oil, and carrying out characteristic fusion to obtain heterogeneous fusion characteristics; a3, performing robustness optimization by adopting a time sequence fluctuation attention suppression mechanism to obtain robust features; a4, respectively calculating the predicted colloid content of the ester oil and the predicted free fatty acid content of the ester oil; a5, calculating the mass prediction fraction of the ester oil; and A6, judging whether the processing result reaches the standard or not based on the predicted colloid content of the ester oil, the predicted free fatty acid content of the ester oil and the predicted mass fraction of the ester oil. The method can solve the problems that raw material characteristic data and processing parameters cannot be deeply fused and the quality of processed raw materials is difficult to pre-judge in advance in a traditional method.

Description

Technical Field

[0001] This invention relates to the field of ester oil technology, and in particular to an intelligent prediction method and system for the quality of ester oil raw materials. Background Technology

[0002] Traditional quality control methods for ester oil raw materials mainly rely on post-processing testing and empirical parameter management. Post-processing testing primarily uses chemical titration and physical separation to determine indicators such as gum content and free fatty acids in the processed raw materials, which makes it difficult to intervene in a timely manner during production. Process management, on the other hand, generally relies solely on fixed pretreatment parameters, such as filtration pressure and temperature, to predict quality. It depends entirely on manual experience to set parameter thresholds and does not take into account the fluctuations in the initial content of gum content and free fatty acids in different batches of ester oil raw materials due to differences in origin, planting environment, and storage conditions.

[0003] With the development of intelligent technology in the oil processing field, existing deep learning technology has been gradually applied to prediction scenarios related to the quality of ester oils. For example, some technologies use a single deep learning model to monitor indicators such as acid value and oxidation degree of ester oils, or combine processing environment parameters to build a simple prediction model. However, these applications still have obvious drawbacks: most existing deep learning models focus on optimizing a single processing parameter or monitoring a single quality indicator for a specific type of ester oil. They fail to deeply integrate and model the core characteristic data of the ester oil raw materials themselves, namely the initial gum content and initial free fatty acid content, with parameters such as filtration pressure and temperature during real-time processing. At the same time, they lack advanced predictive design for the production process and cannot realize the function of predicting the quality of raw materials after processing in advance. This leads to frequent situations in actual production where the pretreatment parameters meet the set standards, but the final product has excessive gum and free fatty acid content due to differences in the initial characteristics of the raw materials. This restricts the refined and efficient control of ester oil raw material processing and makes it difficult to meet the stringent requirements of modern ester oil processing industry for quality stability. Summary of the Invention

[0004] In view of this, the present invention aims to provide an intelligent prediction method and system for the quality of ester oil raw materials, so as to solve the problem that traditional methods cannot deeply integrate raw material characteristic data with processing parameters, making it difficult to predict the quality of raw materials after processing in advance.

[0005] A method for intelligent prediction of ester oil feedstock quality includes:

[0006] A1: Collect base oil characteristic data; collect processing parameter data during the base oil refining process and preprocess it to obtain preprocessed processing parameter data;

[0007] A2: Based on base oil characteristic data and preprocessed processing parameter data, base oil chemical features and processing time sequence features are extracted respectively, and then feature fusion is performed to obtain heterogeneous fusion features; specifically, the base oil chemical features are combined with ester oil characteristic constraint functions to perform feature enhancement processing on the base oil characteristic data, and the processing time sequence features are combined with time sequence continuity constraint functions to capture the time sequence regularity of the preprocessed processing parameter data.

[0008] A3: For the heterogeneous fusion features, a temporal fluctuation suppression attention mechanism is used for robustness optimization to obtain robust features; the temporal fluctuation suppression attention mechanism achieves robust feature generation through the synergistic effect of temporal attention weights and fluctuation suppression coefficients; the fluctuation suppression coefficient is calculated through the statistical characteristics of historical temporal attention weights; the temporal attention weights are obtained based on adaptive learning of heterogeneous fusion features;

[0009] A4: Based on robust characteristics, the predicted content of gum and the predicted content of free fatty acids in ester oil are calculated respectively.

[0010] A5: Calculate the quality prediction score of ester oil based on the predicted content of gum, the predicted content of free fatty acids, and robustness characteristics of ester oil.

[0011] A6: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and the predicted quality score of ester oil, determine whether the processing results meet the standards.

[0012] Furthermore, step A1 also includes:

[0013] A11: Base oil characteristic data are collected by chemical titration device. The data type is numerical data, including gum content and free fatty acid content.

[0014] A12: Processing parameter data during the base oil refining process are collected by pressure and temperature sensors respectively. The data type is time-series numerical data, including filtration pressure and processing temperature. The processing parameter data is processed by moving average method to remove noise and by linear interpolation method to fill in missing values, so as to obtain the cleaned processing parameter data.

[0015] A13: The Z-score standardization method is used to unify the data scale of the cleaned processing parameter data to obtain the preprocessed processing parameter data.

[0016] Furthermore, step A2 also includes:

[0017] A21: Based on base oil characteristic data and pre-processed processing parameter data, the chemical characteristics and processing time sequence characteristics of the base oil are extracted respectively. The calculation method is as follows:

[0018]

[0019]

[0020]

[0021]

[0022] in, Based on the chemical characteristics of base oils, It is a convolutional layer. Based on base oil property data, To sum element by element, For the characteristic constraint function of ester oil, Let t represent the processing time-series characteristics at time t, where t is the time-series index. For gated loop unit, The preprocessed processing parameter data at time t. For Hadama accumulation, For time continuity constraint functions, For batch normalization, For the Sigmoid function, It is a multilayer perceptron. For splicing operations, For transpose operation, The base oil chemical characteristic weight matrix These are the preprocessed machining parameter data at time t-1. For the L2 norm operation, To obtain the maximum value, This refers to the preprocessed processing parameter data. To obtain the minimum value;

[0023] A22: Feature fusion is performed on the chemical characteristics and processing time characteristics of base oils to obtain heterogeneous fused characteristics. The calculation method is as follows:

[0024]

[0025]

[0026] in, Let be the heterogeneous fusion gating weights at time t. Let t represent the heterogeneous fusion characteristics at time t.

[0027] It should be further explained that base oil characteristic data and processing parameter data are heterogeneous types of data. The former is static numerical data, which directly reflects the chemical properties of the raw materials themselves, while the latter is dynamic time-series data, which reflects the evolution of the processing process. At the same time, this scenario has the difficulties of complex attribute relationships between the two types of data, the processing time-series data being easily affected by external interference and resulting in discontinuous fluctuations, and the difficulty in accurately capturing the intrinsic relationship of chemical attributes. This makes it difficult for traditional feature extraction methods to simultaneously adapt to the characteristics of the two types of data and explore their deep relationship with quality indicators.

[0028] The extraction of base oil chemical features is achieved through the synergistic effect of convolutional layers and ester oil characteristic constraint functions. The convolutional layers perform multi-dimensional feature sampling and preliminary screening of base oil characteristic data, completing the transformation from raw data to primary features. The ester oil characteristic constraint function, as a dedicated enhancement module, does not simply modify the primary features, but first deeply mines the base oil characteristic data and its transpose combination information through nonlinear mapping, activating the hidden correlation between different chemical properties. Then, through the feature calibration process, the distribution of the nonlinearly mapped features is adjusted and redundancy is eliminated. At the same time, the intrinsic correlation between gum content and free fatty acid content is strengthened. Through synergistic effect, the correlation between features and ester oil quality indicators is improved, and invalid features are avoided from occupying model resources, so that the extracted base oil chemical features have accurate and efficient characterization capabilities.

[0029] The extraction of processing time-series features relies on the progressive synergy between gated loop units and time-series continuity constraint functions. When processing processing parameter time-series data, the gated loop unit selectively memorizes key time-series patterns from long-term processing, such as the gradual trend of filter pressure and the stable range of processing temperature, while forgetting irrelevant short-term random fluctuations. This effectively captures long-term dependencies in the time-series data, providing a foundation for extracting stable processing patterns. The time-series continuity constraint function focuses on resolving the problem of discontinuous fluctuations in the time-series data. By calculating the correlation between processing parameter data at adjacent time points, it accurately identifies anomalous abrupt changes in the data, such as sudden pressure fluctuations and rapid temperature changes. Non-process interference is mitigated, and the impact of such fluctuations on features is weakened through targeted suppression mechanisms, preventing abnormal interference from obscuring the true temporal patterns of the processing process. The synergy between the two is not a simple functional superposition, but rather the gated loop unit first completes the initial capture of temporal dependencies, and the temporal continuity constraint function then filters out discontinuous fluctuations based on the captured temporal patterns. This retains the core temporal evolution information in the processing parameter data while eliminating invalid fluctuation interference, enabling the extracted processing temporal features to have stable and reliable regularity representation capabilities. This complements the chemical features of the base oil, jointly laying a high-quality feature foundation for the effective fusion of subsequent heterogeneous features.

[0030] Existing technologies for feature extraction of such heterogeneous data typically employ simple feature concatenation or single-network extraction methods. They lack specific enhancement mechanisms tailored to the chemical property correlation patterns of base oil characteristic data and fail to address the fluctuations in processing time-series data. Often, both types of data are directly input into a single model for feature extraction, leading to the loss of crucial chemical property correlation information or interference from fluctuations in the time-series data. In contrast, this invention designs dedicated feature extraction paths and constraint functions for base oil characteristic data and processing parameter data, adapting them to their respective data types and domain characteristics. The extracted base oil chemical features and processing time-series features possess stronger chemical property representation capabilities and stable time-series pattern capture capabilities, effectively avoiding the problems of insufficient generalization, loss of key information, and residual interference components in existing feature extraction technologies.

[0031] Furthermore, the calculation method for the temporal fluctuation suppression attention mechanism in step A3 also includes:

[0032]

[0033]

[0034]

[0035]

[0036]

[0037] in, Let be the temporal attention weights at time t. For the Softmax function, It is a multilayer perceptron. The heterogeneous fusion characteristics at time t are... This represents the initial robustness characteristics at time t. This is a time series fluctuation suppression function. The product is the Hadamard product, and w is the fluctuation suppression coefficient. To sum element by element, For ReLU function, For variance calculation, These are the temporal attention weights at times 1, 2, ..., t, respectively. Let be the robustness feature at time t. For the Laplace operator, This is the robust feature weight matrix.

[0038] It should be further explained that the heterogeneous fusion features simultaneously carry the chemical property information of the base oil and the temporal evolution information of the processing process. The data dimensions are complex and the temporal correlation is strong. However, the equipment noise and process fluctuations that are unavoidable during the processing will cause unstable fluctuation components to be mixed into the heterogeneous fusion features. These fluctuations are not effectively related to the quality indicators of ester oil and are likely to obscure the core laws in the features.

[0039] This invention constructs a temporal fluctuation suppression attention mechanism, addressing the aforementioned problems through temporal attention weights and a temporal fluctuation suppression function. First, the temporal attention weights are generated through deep learning adaptive learning based on heterogeneous fusion features. This enables the identification of core components strongly correlated with ester oil quality indicators within the heterogeneous fusion features, dynamically allocating importance weights for each feature dimension, thus clarifying the focus direction for subsequent optimization. Second, the fluctuation suppression coefficient is calculated by examining the statistical characteristics of historical temporal attention weights. This quantifies the degree of fluctuation in the temporal attention weights, indirectly reflecting the interference intensity of fluctuating components in the heterogeneous fusion features, ensuring the suppression process adapts to the fluctuation characteristics of different processing stages. Finally, the temporal fluctuation suppression function integrates the temporal attention... The weights and fluctuation suppression coefficients are synergistically integrated. Through the linkage between weight allocation and fluctuation suppression, the impact of unstable fluctuations is specifically weakened while highlighting the core feature components, avoiding interference from fluctuation components in the representation of effective information, thus obtaining preliminary robust features. Finally, through the synergistic effect of the robust feature weight matrix and the Laplacian operator, the structural correlation of the preliminary robust features is strengthened, further uncovering the hidden domain structural patterns in the features. Throughout the process, the temporal attention weights provide the focusing direction for fluctuation suppression, the fluctuation suppression coefficients provide the quantitative basis for the suppression strength, and the temporal fluctuation suppression function achieves the synergistic optimization of the two, ensuring that the robust features retain the core information of the heterogeneous fusion features while possessing strong robustness against fluctuation interference.

[0040] Existing technologies for robust optimization of fused features typically employ a single attention mechanism or a simple filtering method. Attention-only schemes focus on the importance allocation of feature dimensions without designing specific suppression mechanisms for temporal fluctuations, making it difficult to effectively filter unstable fluctuation components. Filtering methods, on the other hand, are often static filters with fixed parameters, unable to adjust the suppression intensity according to dynamic changes in features, easily leading to over-filtering of effective temporal information and affecting feature representation capabilities. In contrast, the temporal fluctuation suppression attention mechanism of this invention deeply integrates attention weight allocation and fluctuation suppression. It achieves precise focusing of core features through the attention mechanism and adaptive adjustment of suppression intensity through dynamic calculation of the fluctuation suppression coefficient. Furthermore, it introduces domain-specific constraints for reinforcement, ensuring that the optimized robust features possess both strong representation capabilities and high robustness, better adapting to the actual needs of ester oil quality prediction.

[0041] Furthermore, step A4 also includes:

[0042] A41: Based on robust characteristics, the preliminary predicted content of gum and the preliminary predicted content of free fatty acids in ester oil are calculated respectively. The calculation method is as follows:

[0043]

[0044]

[0045] in, The preliminary predicted gum content of the ester oil at time t. For batch normalization, It is a convolutional layer. This represents the preliminary predicted content of free fatty acids in the ester oil at time t. For the Sigmoid function;

[0046] A42: Constraint calibration is applied to the preliminary predicted content of gum and free fatty acids in ester oils to obtain the predicted content of gum and free fatty acids in ester oils. The calculation method is as follows:

[0047]

[0048]

[0049] in, Let be the predicted gum content of the ester oil at time t. This is a truncation function. These represent the lower limit and upper limit of the predicted gum content for ester oils, respectively. Let be the predicted content of free fatty acids in the ester oil at time t. These represent the lower limit and upper limit of the predicted free fatty acid content of ester oils, respectively. Based on base oil property data, The data represents the preprocessed processing parameters at time t.

[0050] It should be further explained that the core feature of the prediction scenario for gum content and free fatty acid content after ester oil processing is that the formation of these two quality indicators is simultaneously affected by both the chemical properties of the base oil itself and the dynamic parameters of the processing process. Moreover, the correlation between the two and the data shows different patterns. Gum content depends more on the inherent chemical properties of the base oil, while free fatty acid content is more sensitive to the temporal stability of the processing process. Traditional methods cannot deeply integrate raw material characteristic data with processing parameters, resulting in either ignoring the dynamic influence of processing or severing the essential properties of the raw materials, making it difficult to accurately predict the true quality after processing in advance.

[0051] This invention first uses robust features as the core input for prediction, deeply integrating the core chemical attribute information of base oil chemical characteristics with the stable temporal evolution information of processing time series characteristics, breaking down the separation between the two types of data and providing a foundation for the synergistic consideration of dual influencing factors. For gum content prediction, through the synergistic effect of multilayer perceptrons and convolutional layers, it deeply mines the correlation between chemical attributes related to gum formation and the influence of processing in robust features. Then, batch normalization is used to calibrate the feature distribution, avoiding prediction bias caused by data fluctuations and ensuring that the initial prediction can take into account both the raw material nature and processing influences. For free fatty acid content prediction, through time series injection... The synergistic effect of intentional weights and robust features strengthens the dynamic influence of processing time sequence stability on free fatty acid content. Then, nonlinear mapping optimizes the feature characterization to adapt to its sensitivity to processing fluctuations. In the subsequent constraint calibration stage, on the one hand, the preliminary predicted value is limited to a reasonable range in the industry by a truncation function to avoid the prediction results from deviating from the actual production scenario. On the other hand, the ester oil characteristic constraint function and the time sequence continuity constraint function are reused. The former further strengthens the decisive role of raw material chemical properties on gum content, while the latter highlights the key influence of processing parameter time sequence stability on free fatty acid content, ultimately achieving early prediction of post-processing quality indicators.

[0052] Existing technologies for predicting these two types of quality indicators have two main limitations: First, they employ a static, isolated prediction model, calculating solely based on base oil characteristic data using traditional chemical models or simple machine learning models. This ignores the dynamic influence of parameters such as temperature and pressure during processing, resulting in predictions that only reflect the inherent potential of the raw materials and fail to reflect the actual quality after processing, let alone provide advance prediction. Second, they use a shallow, concatenated prediction model, directly concatenating raw material characteristic data and processing parameters into a single model without deep feature fusion and robustness optimization. This leads to insufficient exploration of the intrinsic correlation between the two types of data and a lack of specific prediction logic designed for the differentiated influence patterns of the two types of indicators. This leads to low prediction accuracy, weak anti-interference ability, and difficulty in adapting to changes in different processing scenarios. In contrast, the core advantages of this invention are: First, it achieves deep integration of raw material characteristic data and processing parameters at the feature level. The complete information carried by robust features provides comprehensive support for prediction, solving the pain point of insufficient integration in traditional methods. Second, it designs differentiated prediction paths for the two types of quality indicators, adapting to their respective influencing factors and formation rules, avoiding the problem of insufficient generalization of a single model. Third, it achieves the adaptation of prediction results to actual production scenarios through domain constraint calibration, which not only ensures the feasibility of advance prediction but also improves the prediction accuracy of the prediction results, which is superior to existing technologies.

[0053] Furthermore, the calculation method for the predicted mass fraction of the ester oil in step A5 also includes:

[0054]

[0055]

[0056] in, Let be the weight of the quality index at time t. For splicing operations, These are the benchmark values ​​for the gum content and the free fatty acid content of ester oils, respectively. Let be the predicted mass fraction of the ester oil at time t. Let be the weighted component of the gelatin content in the quality index weights at time t. The free fatty acid content weight component is the weight of the quality index at time t.

[0057] It should be further explained that the predicted quality score for ester oil needs to integrate information on two key quality indicators: gum content and free fatty acid content. Furthermore, the contribution of these two indicators to the overall quality dynamically changes with raw material characteristics and processing conditions. Existing technologies for calculating comprehensive quality scores generally suffer from two limitations: first, static assessment, which uses fixed weights preset by industry experience to simply weight multiple indicators, ignoring the dynamic adjustment of indicator importance caused by differences in raw material characteristics and changes in processing conditions. This makes the score unsuitable for different production scenarios and lacks accuracy; second, information fragmentation, which evaluates based only on raw data of a single indicator or multiple indicators that are not deeply integrated, failing to combine the deep correlation information between raw material characteristics and processing parameters, ignoring the dynamic impact between indicators, resulting in a score that deviates from actual production logic and fails to reflect the overall product quality.

[0058] This invention establishes a method for calculating quality prediction scores. First, the generation of quality indicator weights does not rely on preset fixed values, but is based on the relative ratios of predicted colloid content and predicted free fatty acid content to their corresponding benchmark values. This is achieved through adaptive calculation after normalization, dynamically reflecting the relative performance of the two types of indicators. When a certain indicator deviates further from the benchmark value, its weight component is adjusted accordingly, making the quality prediction score more focused on quality shortcomings and avoiding the average evaluation bias caused by fixed weights. Second, in the core step of score calculation, the predicted contents of the two types of quality are first weighted and integrated with adaptive weight components to reflect their dynamic contribution to overall quality. Then, robust feature deep processing information is incorporated, and the correlation between features and comprehensive quality is further explored through a multilayer perceptron. This ensures that the score not only includes the quantitative performance of explicit indicators but also covers the potential impact of implicit correlations. The resulting quality prediction score accurately quantifies the comprehensive quality level and reflects the true quality status of the processed product in advance, providing a reliable basis for production decisions.

[0059] Furthermore, step A6 also includes:

[0060] A61: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and the predicted quality score of ester oil, and based on industry quality standards, set compliance threshold ranges for gum content, free fatty acid content, and quality prediction score.

[0061] A62: If the predicted content of gum in the ester oil is within the acceptable threshold range of gum content, the predicted content of free fatty acids in the ester oil is within the acceptable threshold range of free fatty acid content, and the predicted quality score of the ester oil is not lower than the acceptable threshold of the predicted quality score, then the processing result is deemed to meet the standard, and the result is fed back to the production control console in real time to keep the current processing flow running continuously.

[0062] A63: If the predicted content of gum in the ester oil exceeds the compliant threshold range, the predicted content of free fatty acids in the ester oil exceeds the compliant threshold range, or the predicted quality score of the ester oil is lower than the qualified threshold, the processing result is deemed substandard, and parameter adjustment suggestions for filtration pressure and processing temperature are generated.

[0063] This invention also discloses an intelligent prediction system for the quality of ester oil raw materials, comprising:

[0064] Base oil data acquisition module: Collects base oil characteristic data; collects processing parameter data during the base oil refining process and preprocesses it to obtain preprocessed processing parameter data;

[0065] Heterogeneous fusion module: Based on base oil characteristic data and preprocessed processing parameter data, the chemical characteristics and processing time sequence characteristics of the base oil are extracted respectively, and then the features are fused to obtain heterogeneous fused features;

[0066] Robustness optimization module: The heterogeneous fusion features are robustly optimized using a temporal fluctuation suppression attention mechanism to obtain robust features;

[0067] Content prediction module: Based on robust features, calculate the predicted content of gum and free fatty acids in ester oil respectively;

[0068] Mass fraction calculation module: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and robustness characteristics, calculate the predicted mass fraction of ester oil.

[0069] Result determination module: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and the predicted quality score of ester oil, determine whether the processing result meets the standard.

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] (1) This invention effectively solves the core problem that traditional methods cannot deeply integrate raw material characteristic data with processing parameters and are difficult to predict the quality of raw materials after processing. By integrating base oil characteristic data and pre-processed processing parameter data, it first extracts and integrates heterogeneous fusion features, and then obtains robust features through robustness optimization. In this way, it accurately predicts the gum content, free fatty acid content and quality prediction score of ester oil, and finally achieves the standard judgment of processing results. It breaks the limitations of traditional post-event testing, fully considers the characteristic differences of different batches of raw materials, and deeply integrates the dynamic law of the processing process. It effectively avoids the quality exceeding the standard problem caused by the initial fluctuation of raw materials or the fluctuation of processing, and provides strong support for the refined and efficient management of ester oil processing.

[0072] (2) In view of the problem that the data types of base oil characteristics data and processing parameters are heterogeneous and the inherent correlation of chemical properties is difficult to capture, which makes it impossible for traditional feature extraction methods to adapt to both types of data characteristics at the same time, this invention innovates the feature extraction scheme: the chemical features of base oil are extracted in collaboration with the convolutional layer and the ester oil characteristic constraint function. After multi-dimensional sampling, nonlinear mapping and feature calibration, the correlation of chemical properties is strengthened; the processing time sequence features rely on the progressive cooperation of the gated loop unit and the time sequence continuity constraint function to capture the long-term dependence of time sequence and filter the interference of discontinuous fluctuations, so that the two types of features have the ability to represent chemical properties and the ability to capture time sequence patterns, respectively, effectively solving the problem of heterogeneous data adaptation and providing a high-quality foundation for subsequent feature fusion.

[0073] (3) In view of the problem that heterogeneous fusion features simultaneously carry chemical properties and temporal evolution information, and that equipment noise and process fluctuations during processing can easily introduce unstable fluctuation components and obscure the core laws, this invention innovatively constructs a temporal fluctuation suppression attention mechanism: by adaptive learning to generate temporal attention weights to identify core components with strong quality correlation, calculating historical weight statistical characteristics to obtain fluctuation suppression coefficients to quantify interference intensity, and then using the temporal fluctuation suppression function to coordinate the two to achieve core feature enhancement and fluctuation weakening, and finally strengthening the feature structure association through robust feature weight matrix and Laplacian operator to accurately solve the fluctuation interference problem, so that the robust features not only completely retain the core information of heterogeneous fusion features, but also have strong anti-interference ability, effectively improving the accuracy and stability of ester oil quality prediction, and adapting to the complex dynamic scenarios of the processing process.

[0074] (4) In view of the problem that the content of gum and free fatty acids are affected by both the chemical properties of raw materials and the dynamic parameters of processing and the correlation patterns are different, this invention uses the robust features that deeply integrate the dual information as the core input to construct a differentiated prediction and constraint calibration system: the gum content prediction is carried out by the collaborative mining of correlation patterns through multilayer perceptron and convolutional layer, and the offset is calibrated by batch normalization; the free fatty acid content prediction is strengthened by the temporal stability influence through temporal attention weight and robust features, and the feature attributes are optimized by nonlinear mapping; then the reasonable range is limited by the truncation function, and the two types of constraint functions are reused to strengthen the specific influencing factors, adapting to the differentiated pattern that the gum content is more dependent on the inherent chemical properties of the base oil and the free fatty acid content is more sensitive to the temporal stability of the processing process, so as to realize the early prediction of the quality indicators after processing.

[0075] (5) In view of the problem that the quality prediction score of ester oil needs to integrate two types of indicators and the contribution changes dynamically with the characteristics of raw materials and processing conditions, this invention constructs a dynamically adapted quality prediction score calculation method: based on the relative ratio of the predicted content of gum and free fatty acids to the benchmark value, an adaptive weight is generated after normalization to dynamically focus on the quality shortcomings; then the weighted and fused explicit indicators and the implicit correlation information of robust features are deeply integrated, and the potential influence is mined through multilayer perceptron, breaking the limitations of static weight and information separation, so that the quality prediction score can accurately quantify the comprehensive quality level, improve the scientificity and practicality of the score, and provide reliable support for production decision-making. Attached Figure Description

[0076] Figure 1 A flowchart illustrating an intelligent prediction method for the quality of ester oil raw materials provided by this invention;

[0077] Figure 2 This is an interface diagram of the result determination module in the ester oil raw material quality prediction system provided by the present invention. Detailed Implementation

[0078] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0079] Example 1: A method for intelligent prediction of ester oil raw material quality, such as... Figure 1 As shown, it includes the following steps:

[0080] A1: Collect base oil characteristic data; collect processing parameter data during the base oil refining process and preprocess it to obtain preprocessed processing parameter data, including:

[0081] A11: Base oil characteristic data are collected by chemical titration device. The data type is numerical data, including gum content and free fatty acid content.

[0082] A12: Processing parameter data during the base oil refining process are collected by pressure and temperature sensors respectively. The data type is time-series numerical data, including filtration pressure and processing temperature. The processing parameter data is processed by moving average method to remove noise and by linear interpolation method to fill in missing values, so as to obtain the cleaned processing parameter data.

[0083] A13: The Z-score standardization method is used to unify the data scale of the cleaned processing parameter data to obtain the preprocessed processing parameter data.

[0084] A2: Based on base oil characteristic data and preprocessed processing parameter data, base oil chemical characteristics and processing time sequence characteristics are extracted respectively, and then feature fusion is performed to obtain heterogeneous fused features, including:

[0085] A21: Based on base oil characteristic data and pre-processed processing parameter data, the chemical characteristics and processing time sequence characteristics of the base oil are extracted respectively. The calculation method is as follows:

[0086]

[0087]

[0088]

[0089]

[0090] in, Based on the chemical characteristics of base oils, It is a convolutional layer. Based on base oil property data, To sum element by element, For the characteristic constraint function of ester oil, Let t represent the processing time-series characteristics at time t, where t is the time-series index. For gated loop unit, The preprocessed processing parameter data at time t. For Hadama accumulation, For time continuity constraint functions, For batch normalization, For the Sigmoid function, It is a multilayer perceptron. For splicing operations, For transpose operation, The base oil chemical characteristic weight matrix These are the preprocessed machining parameter data at time t-1. For the L2 norm operation, To obtain the maximum value, This refers to the preprocessed processing parameter data. To obtain the minimum value;

[0091] A22: Feature fusion is performed on the chemical characteristics and processing time characteristics of base oils to obtain heterogeneous fused characteristics. The calculation method is as follows:

[0092]

[0093]

[0094] in, Let be the heterogeneous fusion gating weights at time t. Let t represent the heterogeneous fusion characteristics at time t.

[0095] Specifically, for scenarios where the chemical characteristics of base oils fluctuate significantly due to batch-to-batch variations in raw materials during ester oil processing, and where processing sequence characteristics exhibit long-term time dependence, this invention also provides a heterogeneous fusion calculation method with time-decay residuals to replace step A22. The calculation method is as follows:

[0096]

[0097]

[0098]

[0099] in, The time decay coefficient at time t, It is an exponential function. This refers to the total processing time for a single batch. The start time of processing Calculated as the mean.

[0100] Both multilayer perceptrons in the ester oil characteristic constraint function adopt a 3-layer fully connected structure: the first layer has 64 neurons, the second layer has 32 neurons, and the third layer has 16 neurons. The activation function is Sigmoid. The dimension of the base oil chemical feature weight matrix is ​​set to 16×16 to match the dimension of the MLP output features.

[0101] The gated recurrent unit is set to a 2-layer stacked structure, with 64 neurons in each hidden layer, a dropout rate of 0.2, an input sequence length of 30 sampling points, and an output temporal feature vector of 64 dimensions. The reset gate and update gate of the gated recurrent unit both use the Sigmoid function as the activation function, the candidate state activation function uses the hyperbolic tangent function, and the gradient clipping threshold is set to 1.0.

[0102] A3: For the heterogeneous fusion features, a temporal fluctuation suppression attention mechanism is used for robustness optimization to obtain robust features. The calculation method of the temporal fluctuation suppression attention mechanism in step A3 includes:

[0103]

[0104]

[0105]

[0106]

[0107]

[0108] in, Let be the temporal attention weights at time t. For the Softmax function, It is a multilayer perceptron. The heterogeneous fusion characteristics at time t are... This represents the initial robustness characteristics at time t. This is a time series fluctuation suppression function. The product is the Hadamard product, and w is the fluctuation suppression coefficient. To sum element by element, For ReLU function, For variance calculation, These are the temporal attention weights at times 1, 2, ..., t, respectively. Let be the robustness feature at time t. For the Laplace operator, This is the robust feature weight matrix.

[0109] A4: Based on robust characteristics, the predicted content of gum and the predicted content of free fatty acids in ester oils are calculated, including:

[0110] A41: Based on robust characteristics, the preliminary predicted content of gum and the preliminary predicted content of free fatty acids in ester oil are calculated respectively. The calculation method is as follows:

[0111]

[0112]

[0113] in, The preliminary predicted gum content of the ester oil at time t. For batch normalization, It is a convolutional layer. This represents the preliminary predicted content of free fatty acids in the ester oil at time t. For the Sigmoid function;

[0114] A42: Constraint calibration is applied to the preliminary predicted content of gum and free fatty acids in ester oils to obtain the predicted content of gum and free fatty acids in ester oils. The calculation method is as follows:

[0115]

[0116]

[0117] in, Let be the predicted gum content of the ester oil at time t. This is a truncation function. These represent the lower limit and upper limit of the predicted gum content for ester oils, respectively. Let be the predicted content of free fatty acids in the ester oil at time t. These represent the lower limit and upper limit of the predicted free fatty acid content of ester oils, respectively. Based on base oil property data, The data represents the preprocessed processing parameters at time t.

[0118] A5: Based on the predicted content of gums, the predicted content of free fatty acids, and robustness characteristics of the ester oil, calculate the predicted mass score of the ester oil. The calculation method for the predicted mass score of the ester oil in step A5 includes:

[0119]

[0120]

[0121] in, Let be the weight of the quality index at time t. For splicing operations, These are the benchmark values ​​for the gum content and the free fatty acid content of ester oils, respectively. Let be the predicted mass fraction of the ester oil at time t. Let be the weighted component of the gelatin content in the quality index weights at time t. The free fatty acid content weight component is the weight of the quality index at time t.

[0122] A6: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and the predicted quality score of ester oil, determine whether the processing results meet the standards, including:

[0123] A61: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and the predicted quality score of ester oil, and based on industry quality standards, set compliance threshold ranges for gum content, free fatty acid content, and quality prediction score.

[0124] A62: If the predicted content of gum in the ester oil is within the acceptable threshold range of gum content, the predicted content of free fatty acids in the ester oil is within the acceptable threshold range of free fatty acid content, and the predicted quality score of the ester oil is not lower than the acceptable threshold of the predicted quality score, then the processing result is deemed to meet the standard, and the result is fed back to the production control console in real time to keep the current processing flow running continuously.

[0125] A63: If the predicted content of gum in the ester oil exceeds the compliant threshold range, the predicted content of free fatty acids in the ester oil exceeds the compliant threshold range, or the predicted quality score of the ester oil is lower than the qualified threshold, the processing result is deemed substandard, and parameter adjustment suggestions for filtration pressure and processing temperature are generated.

[0126] For example, the compliance thresholds set in step A61 are as follows:

[0127] The compliant threshold range for gum content is ≤0.05% (mass fraction).

[0128] The compliant threshold range for free fatty acid content is ≤0.03% (mass fraction).

[0129] Quality prediction score passing threshold: ≥90 points (out of 100 points, the higher the score, the better the overall quality).

[0130] For a specific batch of ester oil base oil refining process, the following calculations were performed using the method of this invention:

[0131] Predicted gum content of ester oil: 0.052% (slightly exceeding the upper limit of the compliance threshold);

[0132] Predicted free fatty acid content of ester oil: 0.028% (within the compliance threshold range);

[0133] Predicted quality score for ester oil: 88 points (slightly below standard).

[0134] Because the predicted gum content exceeds the threshold and the predicted quality score is below the acceptable threshold, the processing result of this batch is deemed substandard at this time. Figure 2 As shown;

[0135] A rule base was constructed for filtration pressure and processing temperature. The rule dimensions cover the correlation between the degree of excessive colloid content, the extent of low mass fraction, and processing parameters. Specific rules and adjustment suggestions are shown in Table 1.

[0136] Table 1. Rules for Adjusting Ester Oil Filtration Pressure and Processing Temperature

[0137] Rule Number Triggering conditions Filter pressure adjustment rules Processing temperature adjustment rules R1 The predicted content of colloids exceeds the upper limit by 0-0.005%, the predicted content of free fatty acids meets the standard, and the mass fraction is 0-5 points lower. Reduce the filtration pressure by 3% to 5% of the current value. Reduce the processing temperature by 1% to 2% of the current value. R2 The predicted content of colloids exceeds the upper limit by more than 0.005%, the predicted content of free fatty acids meets the standard, but the mass fraction is 0-5 points lower. Reduce the filtration pressure by 5% to 8% of the current value. Reduce the processing temperature by 2% to 3% of the current value. R3 The predicted content of colloids meets the standard, but the predicted content of free fatty acids exceeds the upper limit, and the mass score is 0-5 points lower. Increase the filtration pressure by 2% to 4% of the current value. Increase the filtration pressure by 2% to 4% of the current value.

[0138] This trigger rule R1 (colloidal content exceeds the upper limit by 0.002%, predicted free fatty acid content meets the standard, mass fraction is 2 points low), combined with the current processing parameters of this batch (filtration pressure: 0.8MPa, processing temperature: 85℃), generates specific adjustment suggestions:

[0139] Filtration pressure adjustment: Reduced by 3% from the current 0.8MPa to 0.77MPa;

[0140] Processing temperature adjustment: Reduced by 1.5% from the current 85℃ to 83℃.

[0141] The neural network modules involved in this invention, such as convolutional layers, gated recurrent units, multilayer perceptrons, and time-series fluctuation suppression attention mechanisms, employ existing end-to-end training methods for model training and parameter optimization. Specifically, based on preprocessed base oil characteristic data, processing parameter data, and corresponding actual quality index data, training, validation, and test sets are constructed and divided in a 7:1:2 ratio. Mean squared error is used as the core loss function, combined with a multi-task weighted loss strategy to balance the prediction errors of gum content, free fatty acid content, and quality score. The Adam optimizer is used to iterate over all learnable parameters of the model. The updated system employs a tiered learning rate scheduling strategy to dynamically adjust the learning rate and improve convergence efficiency. During training, a batch training mode is used, with Dropout regularization to suppress overfitting. An early stopping strategy is introduced, determining the termination time based on the validation set prediction accuracy reaching a set value to avoid overtraining. In the model initialization phase, the Xavier initialization method is used to initialize the weight parameters of modules such as convolutional layers and multilayer perceptrons, ensuring stable convergence during training. The training processes of all modules are unified and interconnected, with backpropagation algorithm used to achieve collaborative optimization of parameters across the entire chain, ultimately resulting in a stable and accurate model.

[0142] Example 2: This invention also discloses an intelligent prediction system for the quality of ester oil raw materials, comprising:

[0143] Base oil data acquisition module: Collects base oil characteristic data; collects processing parameter data during the base oil refining process and preprocesses it to obtain preprocessed processing parameter data;

[0144] Heterogeneous fusion module: Based on base oil characteristic data and preprocessed processing parameter data, the chemical characteristics and processing time sequence characteristics of the base oil are extracted respectively, and then the features are fused to obtain heterogeneous fused features;

[0145] Robustness optimization module: The heterogeneous fusion features are robustly optimized using a temporal fluctuation suppression attention mechanism to obtain robust features;

[0146] Content prediction module: Based on robust features, calculate the predicted content of gum and free fatty acids in ester oil respectively;

[0147] Mass fraction calculation module: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and robustness characteristics, calculate the predicted mass fraction of ester oil.

[0148] Result determination module: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and the predicted quality score of ester oil, determine whether the processing result meets the standard.

[0149] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0150] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0151] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for intelligent prediction of the quality of ester oil raw materials, characterized in that, Includes the following steps: A1: Collect base oil characteristic data; collect processing parameter data during the base oil refining process and preprocess it to obtain preprocessed processing parameter data; A2: Based on base oil characteristic data and preprocessed processing parameter data, base oil chemical features and processing time sequence features are extracted respectively, and then feature fusion is performed to obtain heterogeneous fusion features; specifically, the base oil chemical features are combined with ester oil characteristic constraint functions to perform feature enhancement processing on the base oil characteristic data, and the processing time sequence features are combined with time sequence continuity constraint functions to capture the time sequence regularity of the preprocessed processing parameter data. A3: Robust features are obtained by using a temporal fluctuation suppression attention mechanism to optimize the heterogeneous fusion features. The temporal fluctuation suppression attention mechanism achieves robust feature generation through the synergistic effect of temporal attention weights and fluctuation suppression coefficients; the fluctuation suppression coefficients are calculated using the statistical characteristics of historical temporal attention weights; and the temporal attention weights are obtained through adaptive learning based on heterogeneous fusion features. A4: Based on robust characteristics, the predicted content of gum and the predicted content of free fatty acids in ester oil are calculated respectively. A5: Calculate the quality prediction score of ester oil based on the predicted content of gum, the predicted content of free fatty acids, and robustness characteristics of ester oil. A6: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and the predicted quality score of ester oil, determine whether the processing results meet the standards.

2. The intelligent prediction method for ester oil raw material quality according to claim 1, characterized in that, Step A1 includes: A11: Base oil characteristic data are collected by chemical titration device. The data type is numerical data, including gum content and free fatty acid content. A12: Processing parameter data during the base oil refining process are collected by pressure and temperature sensors respectively. The data type is time-series numerical data, including filtration pressure and processing temperature. The processing parameter data is processed by moving average method to remove noise and by linear interpolation method to fill in missing values, so as to obtain the cleaned processing parameter data. A13: The Z-score standardization method is used to unify the data scale of the cleaned processing parameter data to obtain the preprocessed processing parameter data.

3. The intelligent prediction method for ester oil raw material quality according to claim 2, characterized in that, Step A2 includes: A21: Based on base oil characteristic data and preprocessed processing parameter data, extract the base oil chemical characteristics and processing time sequence characteristics respectively, calculated as follows: in, Based on the chemical characteristics of base oils, It is a convolutional layer. Based on base oil property data, To sum element by element, For the characteristic constraint function of ester oil, Let t represent the processing time-series characteristics at time t, where t is the time-series index. For gated loop unit, The preprocessed processing parameter data at time t. For Hadama accumulation, For time continuity constraint functions, For batch normalization, For the Sigmoid function, It is a multilayer perceptron. For splicing operations, For transpose operation, The chemical characteristic weight matrix of base oil These are the preprocessed machining parameter data at time t-1. For the L2 norm operation, To obtain the maximum value, This is the preprocessed processing parameter data. To obtain the minimum value; A22: Feature fusion is performed on the chemical characteristics and processing time characteristics of base oils to obtain heterogeneous fused characteristics. The calculation method is as follows: in, Let be the heterogeneous fusion gating weights at time t. The heterogeneous fusion feature at time t.

4. The intelligent prediction method for ester oil raw material quality according to claim 1, characterized in that, The calculation method for the temporal fluctuation suppression attention mechanism in step A3 includes: in, Let be the temporal attention weights at time t. For the Softmax function, It is a multilayer perceptron. The heterogeneous fusion characteristics at time t are... The initial robustness characteristics at time t. This is a time series fluctuation suppression function. The product is the Hadamard product, and w is the fluctuation suppression coefficient. To sum element by element, For ReLU function, For variance calculation, These are the temporal attention weights at times 1, 2, ..., t, respectively. Let be the robustness feature at time t. For the Laplace operator, This is the robust feature weight matrix.

5. The intelligent prediction method for ester oil raw material quality according to claim 4, characterized in that, The A4 step includes: A41: Based on robust characteristics, the preliminary predicted content of gum and the preliminary predicted content of free fatty acids in ester oil are calculated respectively. The calculation method is as follows: in, The preliminary predicted gum content of the ester oil at time t. For batch normalization, It is a convolutional layer. This represents the preliminary predicted content of free fatty acids in the ester oil at time t. For the Sigmoid function; A42: Constraint calibration is applied to the preliminary predicted content of gum and free fatty acids in ester oils to obtain the predicted content of gum and free fatty acids in ester oils. The calculation method is as follows: in, Let be the predicted gum content of the ester oil at time t. This is a truncation function. These represent the lower limit and upper limit of the predicted gum content for ester oils, respectively. Let be the predicted content of free fatty acids in the ester oil at time t. These represent the lower limit and upper limit of the predicted free fatty acid content of ester oils, respectively. Based on base oil property data, The preprocessed processing parameter data at time t. For the characteristic constraint function of ester oil, This is the temporal continuity constraint function.

6. The intelligent prediction method for ester oil raw material quality according to claim 5, characterized in that, The method for calculating the predicted mass fraction of the ester oil in step A5 includes: in, Let be the weight of the quality index at time t. For splicing operations, These are the benchmark values ​​for the gum content and the free fatty acid content of ester oils, respectively. Let be the predicted mass fraction of the ester oil at time t. Let be the weighted component of the gelatin content in the quality index weights at time t. The free fatty acid content weight component is the weight of the quality index at time t.

7. The intelligent prediction method for ester oil raw material quality according to claim 6, characterized in that, Step A6 includes: A61: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and the predicted quality score of ester oil, and based on industry quality standards, set compliance threshold ranges for gum content, free fatty acid content, and quality prediction score. A62: If the predicted content of gum in the ester oil is within the acceptable threshold range of gum content, the predicted content of free fatty acids in the ester oil is within the acceptable threshold range of free fatty acid content, and the predicted quality score of the ester oil is not lower than the acceptable threshold of the predicted quality score, then the processing result is deemed to meet the standard, and the result is fed back to the production control console in real time to keep the current processing flow running continuously. A63: If the predicted content of gum in the ester oil exceeds the compliant threshold range, the predicted content of free fatty acids in the ester oil exceeds the compliant threshold range, or the predicted quality score of the ester oil is lower than the qualified threshold, the processing result is deemed substandard, and parameter adjustment suggestions for filtration pressure and processing temperature are generated.

8. A smart prediction system for the quality of ester oil raw materials, characterized in that, include: Base oil data acquisition module: Collects base oil characteristic data; collects processing parameter data during the base oil refining process and preprocesses it to obtain preprocessed processing parameter data; Heterogeneous fusion module: Based on base oil characteristic data and preprocessed processing parameter data, the chemical characteristics and processing time sequence characteristics of the base oil are extracted respectively, and then the features are fused to obtain heterogeneous fused features; Robustness optimization module: The heterogeneous fusion features are robustly optimized using a temporal fluctuation suppression attention mechanism to obtain robust features; Content prediction module: Based on robust features, calculate the predicted content of gum and free fatty acids in ester oil respectively; Mass fraction calculation module: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and robustness characteristics, calculate the predicted mass fraction of ester oil. Result determination module: Based on the predicted content of gum in ester oil, the predicted content of free fatty acids in ester oil, and the predicted quality score of ester oil, determine whether the processing result meets the standard; so as to realize the intelligent prediction method for the quality of ester oil raw materials as described in any one of claims 1-7.

Images